基于条件互信息的低冗余短期负荷预测特征选择

Low Redundancy Feature Selection Using Conditional Mutual Information for Short-Term Load Forecasting

为避免负荷预测特征集中冗余特征对预测精度的负面影响,降低预测器复杂度,提出一种基于条件互信息(CMI)和高斯过程回归(GPR)的短期负荷预测特征选择方法.首先,为降低建模所用特征量,根据与目标变量具有最大互信息的特征,选取剩余特征中可对目标变量提供最大信息增益的特征,计算CMI值并进行排序;然后,以GPR为预测器,以其预测结果平均绝对百分比误差为决策变量,按照特征CMI值排序顺序,采用序列前向选择方法,确定最优特征子集;最终,以最优特征子集构建GPR预测模型,并与皮尔逊相关系数法(PCC)和互信息(MI)2种特征选择方法分别结合支持向量机和反向传播神经网络开展对比实验.实验结果证明新方法降低了最优特征集合冗余度与预测模型复杂度,且具有更高的预测精度.

In order to avoid the influence of redundancy features which will affect the predict accuracy of load forecasting and reduce the complexity of predictor,a method based on Gaussian Process Regression (GPR) and Conditional Mutual Information (CMI) for feature selection is proposed for short-term load forecasting.Firstly,to decrease the number of features when building a forecasting model,a feature is selected as the one with the highest mutual information (MI) with the target variable.Then,respect to the selected feature,the next feature which adds the largest information to the target variable is selected.The features are ranked in a descending order according to calculating the CMI.Secondly,according to the order of CMI,a sequential forward selection method with Mean Absolute Percentage Error of GPR is utilized for choosing the optimal feature subset.Finally,GPR with the optimal subset is used for building the forecasting model for load prediction,and compared with support vector machine and back propagation neural network combined with Pearson Correlation Coefficient (PCC) and MI respectively.The results show that the new method reduces the redundancy of the optimal feature set and has a higher forecast accuracy with lower complexity of structure.